{"paper_id":"2448938e-2c44-4e86-bafe-f35d851ab2bf","body_text":"Dual Architecture of Consciousness Transitions: Evidence from Sleep EEG \nAnalysis \nEmma Dobbin \nConsciousness Gradient Theory Group Ltd, United Kingdom \nAbstract \nThe neural mechanisms underlying consciousness transitions have remained contentious, with \nIntegrated Information Theory (IIT) predicting gradual changes and Global Workspace Theory \n(GWT) predicting discrete threshold events. Using an adaptive consciousnes s measurement \nframework applied to 622 sleep stage transitions across 12 subjects from the Sleep -EDF \ndatabase, we demonstrate that both theories are correct within distinct transition types. Our \nConsciousness Gradient Index (CGI) framework reveals that thr ee of four transition types \n(75%: Wake→N1, N2→N3, REM→Wake) follow gradual dynamics consistent with IIT \npredictions (mean slopes: -0.736, -0.532, +0.614), while one transition type (25%: N1→N2) \nexhibits threshold behavior via thalamic spindle gating mechanisms (mean slope: -2.188). The \nN1→N2 transition showed significantly steeper CGI decline compared to Wake→N1 (paired \nt-test: t= -9.334, p<0.001, Cohen's d= -2.814) and N2→N3 (t= -6.341, p<0.001, d= -1.912), \nsupporting a dual architecture model where consciousness transitions employ both continuous \nintegration mechanisms and discrete neural gates. These findings reconcile competing \ntheoretical frameworks and establish a unified model for understanding consciousness state \nchanges across biological systems. \nKeywords: consciousness measurement, sleep stages, EEG analysis, EEG complexity, sleep \nonset, spindle gating, threshold gating, Integrated Information Theory, Global Workspace \nTheory, dual architecture \nIntroduction \nThe measurement and understanding of consciousness state changes represents a fundamental \nchallenge in neuroscience and philosophy of mind. Two major theoretical frameworks have \ndominated recent discourse: Integrated Information Theory (IIT), which predict s gradual \nconsciousness changes through continuous information integration (Tononi et al., 2016), and \nGlobal Workspace Theory (GWT), which proposes discrete broadcasting events and threshold \nmechanisms (Dehaene & Changeux, 2011). These frameworks have been treated as competing \nexplanations, yet empirical evidence supporting each perspective has remained inconclusive. \nSleep provides an ideal natural laboratory for investigating consciousness transitions, as \nindividuals cycle predictably through distinct states with well-characterized neural signatures. \nTraditional approaches to measuring consciousness during sleep have relied on subjective self-\nreports or binary state classifications (awake vs. asleep), which fail to capture the continuous \nnature of consciousness gradients. More recent quantitative approaches have attempted to \nmeasure consciousness using complexity metrics (Casali et al., 2013) or connectivity measures \n(Tagliazucchi et al., 2016), but these methods have not systematically evaluated whether \ndifferent transitions follow distinct mechanistic patterns. \nWe developed the Consciousness Gradient Index (CGI), an adaptive measurement framework \nthat employs state-specific neural metrics to quantify consciousness level changes during sleep \ntransitions. The formula CGI = √(φ × ρ) × 10 combines information integra tion (φ) with \nadaptive response capacity (ρ), where φ is calculated using transition -appropriate neural \nfeatures: alpha power (8-13 Hz) for arousal transitions, inverted spindle density (11-16 Hz) for \nthe N1→N2 threshold, and spectral entropy for deep sleep transitions. This adaptive approach \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted November 11, 2025. ; https://doi.org/10.1101/2025.11.10.687628doi: bioRxiv preprint \n\nallows the framework to capture both gradual and discrete consciousness changes using the \nsame mathematical structure. \nHere we report the results of applying this framework to 622 sleep stage transitions from 12 \nhealthy adults, revealing a dual architecture where distinct neural mechanisms govern different \ntransition types. We demonstrate that the apparent conflict between IIT and GWT reflects their \napplication to different classes of consciousness transitions, and propose a unified model \nreconciling these theoretical perspectives. \nMethods \nParticipants and Data \nWe analyzed  polysomnographic recordings from 12 healthy adults (6 male, 6 female, age \nrange: 25 -35 years, mean age: 29.3 ± 3.2 years) from the Sleep -EDF Database Expanded \n(Kemp et al., 2000). The database contains whole-night recordings with sampling frequency of \n100 Hz. All recordings were obtained from participants with no reported sleep disorders, \nneurological conditions, or psychoactive medication use. Ethical approval for the original study \nwas obtained by the data providers, and the database is publicly available for research purposes. \nEEG Recording and Preprocessing \nTwo EEG channels (Fpz-Cz and Pz-Oz) were used for all analyses to ensure signal specificity \nto cortical activity and minimize contamination from respiratory or muscular artifacts. Raw \nEEG signals were bandpass filtered (0.5-30 Hz) using a zero-phase Butterworth filter to remove \ndrift and high -frequency noise. Sleep stages were scored in 30 -second epochs according to \nRechtschaffen and Kales criteria (Rechtschaffen & Kales, 1968), which maps equivalently to \nAASM N1 -N3 stages for the purposes of this analysis.  Sleep scoring was performed by \ncertified sleep technicians. Hypnograms were provided with the database and used to identify \nstate transitions. \nTransition Detection and Analysis \nWe focused on four key transition types representing different consciousness change \nmechanisms: Wake→N1 (sleep onset), N1→N2 (spindle emergence), N2→N3 (deep sleep \nentry), and REM→Wake (awakening). For each detected transition, we extracted EEG data \nfrom 150 seconds before to 150 seconds after the transition point (total window: 300 seconds). \nCGI values were calculated in 30 -second sliding epochs with 15 -second overlap, yielding 19 \nCGI measurements per transition. Linear slopes were fit to CGI values across  time using \nordinary least squares regression to quantify the rate of consciousness change. \nAdaptive CGI Calculation \nThe core formula CGI = √( φ × ρ) × 10 was applied to all transitions, with ρ fixed at 1.0 for \nEEG analysis. The critical innovation lies in the adaptive calculation of φ (information \nintegration) based on transition-specific neural mechanisms: \n• Wake→N1 and REM→Wake transitions: φ calculated from alpha power (8-13 Hz), \nreflecting arousal gradient mechanisms. Power spectral density was computed using \nWelch's method with 2-second windows and 50% overlap. \n• N1→N2 transition: φ calculated from inverted spindle density (11-16 Hz), reflecting \nthreshold gating mechanisms. Spindles were detected using RMS power exceeding \nmean + 2 standard deviations with duration criteria of 0.5-2 seconds. The inversion (φ \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted November 11, 2025. ; https://doi.org/10.1101/2025.11.10.687628doi: bioRxiv preprint \n\n= 1 - normalized_spindle_density) captures consciousness suppression via thalamic \ngating. \n• N2→N3 transition: φ calculated from spectral entropy (0.5-30 Hz), reflecting \ncomplexity reduction. Entropy was computed from the probability distribution of \nnormalized power spectral density across frequency bins. \nAll φ values were normalized to a 0 -10 scale before CGI calculation to ensure comparability \nacross metrics. \nStatistical Analysis \nTo address the repeated measures structure of our data (multiple transitions per subject), we \nemployed a conservative subject-level analysis approach. For each subject and transition type, \nwe calculated the mean slope across all detected transitions. Statistical comparisons were then \nperformed on these subject -level means using paired t -tests (for transitions present in all \nsubjects) and repeated measures ANOVA (for overall differences). This approach \nappropriately accounts for within -subject dependencies w hile avoiding pseudoreplication. \nEffect sizes were calculated using Cohen's d for paired comparisons. Statistical significance \nwas set at p < 0.05 (two-tailed). Variance decomposition was performed to quantify between-\nsubject versus within-subject variability using intraclass correlation coefficients (ICC). \nResults \nDataset Characteristics \nAcross 12 subjects, we analyzed  622 total sleep stage transitions: 136 Wake→N1 transitions \n(mean 11.3 per subject), 207 N1→N2 transitions (17.2 per subject), 247 N2→N3 transitions \n(20.6 per subject), and 32 REM→Wake transitions (4.6 per subject in 7 of 12 subjects). The \nlower frequency of REM→Wake transitions reflects both the typical sleep architecture pattern \nand our conservative transition detection criteria. \nSubject-Level Slope Characteristics \nAnalysis at the subject level revealed distinct patterns across transition types: \n• Wake→N1: Mean slope = -0.736 ± 0.266 (SD), n = 12 subjects, SE = 0.077 \n• N1→N2: Mean slope = -2.188 ± 0.545 (SD), n = 12 subjects, SE = 0.157 \n• N2→N3: Mean slope = -0.532 ± 0.526 (SD), n = 12 subjects, SE = 0.152 \n• REM→Wake: Mean slope = +0.614 ± 1.080 (SD), n = 7 subjects, SE = 0.408 \nThe N1→N2 transition showed substantially steeper slopes (approximately 3 -fold greater \nmagnitude) compared to other NREM transitions, consistent with a threshold mechanism. \nREM→Wake was the only transition showing positive slopes, reflecting consciousness \nrestoration during awakening. Complete subject-level statistics are presented in Table 1. \n \nTable 1. Subject-level mean slopes for each transition type \nTransition Type N Subjects Mean Slope SD 95% CI \nWake→N1 12 -0.736 0.266 [-0.900, -0.573] \nN1→N2 12 -2.188 0.545 [-2.532, -1.844] \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted November 11, 2025. ; https://doi.org/10.1101/2025.11.10.687628doi: bioRxiv preprint \n\nTransition Type N Subjects Mean Slope SD 95% CI \nN2→N3 12 -0.532 0.526 [-0.876, -0.188] \nREM→Wake 7 +0.614 1.080 [-0.393, +1.621] \nNote: CI = Confidence Interval. N1→N2 mean slope is bolded to highlight the steepest decline.  \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted November 11, 2025. ; https://doi.org/10.1101/2025.11.10.687628doi: bioRxiv preprint \n\n \n \n \nStatistical Comparison of Transition Types \nRepeated measures ANOVA on subject -level means for the three main NREM transitions \n(Wake→N1, N1→N2, N2→N3) revealed highly significant differences (F = 45.564, p < \n0.001), confirming that transition types differ fundamentally in their consciousness change \ndynamics. \nPaired t-tests comparing N1→N2 against other transitions demonstrated: \n• N1→N2 vs. Wake→N1: t(11) = -9.334, p < 0.001, Cohen's d = -2.814 (very large \neffect) \n• N1→N2 vs. N2→N3: t(11) = -6.341, p < 0.001, Cohen's d = -1.912 (very large effect) \nThese extremely large effect sizes indicate that the N1→N2 transition operates through a \nqualitatively different mechanism than other sleep transitions. Complete pairwise comparison \nstatistics are presented in Table 2. \n \nTable 2. Pairwise comparisons of transition types \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted November 11, 2025. ; https://doi.org/10.1101/2025.11.10.687628doi: bioRxiv preprint \n\nComparison t-statistic p-value Cohen's d \nN1→N2 vs Wake→N1 -9.334 <0.001 -2.814 \nN1→N2 vs N2→N3 -6.341 <0.001 -1.912 \nNote: All comparisons are paired t -tests (n=12 subjects for NREM transitions). p -values <0.001 are bolded to \nhighlight statistical significance. \nVariance Structure Analysis \nFor the N1→N2 transition, variance decomposition revealed that 96.6% of total variance was \nwithin-subject (across individual transitions) while only 3.6% was between -subject (across \nindividuals). The intraclass correlation coefficient (ICC = 0.036) indicat ed low clustering, \nmeaning the threshold mechanism operates consistently across individuals despite high \nvariability in timing of individual spindle events. Notably, 15% of individual N1→N2 \ntransitions showed near-zero slopes (|slope| < 0.5), representing cases where spindle onset was \ngradual rather than abrupt. This biological variability does not undermine the overall pattern, \nas the subject -level analysis demonstrates consistent directional effects. A supplemental \nrandom-intercept mixed effects model con firmed that N1→N2 transition type contributed an \nadditional -1.457 CGI decline per second (95% CI: [-1.891, -1.023], p < 0.001) beyond subject-\nlevel random effects, providing independent confirmation of the threshold mechanism. \nDiscussion \nDual Architecture Model \nOur findings establish that consciousness transitions employ two distinct architectural \nprinciples operating in parallel across the sleep cycle. The majority of transitions (75%: \nWake→N1, N2→N3, REM→Wake) follow gradual dynamics consistent with Integrated \nInformation Theory's prediction of continuous φ modulation through information integration \nchanges. However, one critical transition (25%: N1→N2) exhibits threshold behavior mediated \nby thalamic spindle gating, consistent with Global Workspace Theory's dis crete broadcasting \nmechanism. This dual architecture reconciles the apparent contradiction between these major \ntheoretical frameworks by demonstrating that both correctly describe consciousness \ntransitions—but in different contexts. \nSpindle Gating as Threshold Mechanism \nThe N1→N2 transition's distinctive pattern emerges from the functional role of sleep spindles, \nwhich are generated by thalamic reticular nucleus (TRN) neurons and propagate through \nthalamocortical circuits (Lüthi, 2014). Spindles actively suppress sensory processing and \ncortical communication (Dang -Vu et al., 2010), creating a neural \"gate\" that discretely \ntransitions the brain from light sleep (where external stimuli can still penetrate) to stable Stage \n2 sleep (where consciousness of external environment is largely abolished). Our demonstration \nthat inverted spindle density provides the appropriate φ metric captures this gating function: \nincreasing spindle activity directly reduces information integration, producing the steep CGI \ndecline we observed. Supplemental analysis using a random-intercept mixed model confirmed \nan additional -1.457 CGI decline specifically attributable to N1→N2 transition type (p < \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted November 11, 2025. ; https://doi.org/10.1101/2025.11.10.687628doi: bioRxiv preprint \n\n0.001), beyond subject -level variation, further supporting the threshold mechanism \ninterpretation. \nThe high within -subject variability in N1→N2 slopes (96.6% of variance) reflects the \nstochastic nature of spindle generation. While the overall direction is consistent (consciousness \ndecline), the precise timing and abruptness of spindle onset varies acros s individual transition \nevents. This biological variability is entirely consistent with threshold mechanisms in neural \nsystems, which exhibit probabilistic rather than deterministic triggering (Poulet & Petersen, \n2008). \nGradual Transitions via Integration Modulation \nIn contrast to the spindle-gated N1→N2 transition, the Wake→N1, N2→N3, and REM→Wake \ntransitions all showed relatively gradual slopes reflecting continuous modulation of \ninformation integration. The Wake→N1 transition (mean slope -0.736) captures the \nprogressive reduction in arousal systems mediated by descending noradrenergic and \ncholinergic activity (Brown et al., 2012). The N2→N3 transition (mean slope -0.532) reflects \ngradual increases in slow wave activity and cortical bistability as homeostatic sleep p ressure \naccumulates (Vyazovskiy et al., 2009). The REM→Wake transition (+0.614) involves \nprogressive reactivation of ascending arousal systems and cholinergic drive (Brown et al., \n2012). \nThese gradual transitions align with IIT's prediction that consciousness level scales with \nintegrated information. Changes in global brain state (arousal level, slow wave proportion, \ncomplexity) produce corresponding changes in φ through their effects on effective connectivity \nand information integration capacity (Tononi et al., 2016). Importantly, none of these \ntransitions involve discrete gating mechanisms analogous to spindle emergence. \nMethodological Considerations \nOur approach addresses several critical methodological challenges in consciousness \nmeasurement. First, the adaptive φ calculation ensures that different transitions are measured \nusing their mechanistically appropriate neural features rather than forcing al l transitions \nthrough the same metric. This biological validity enhances the framework's ability to detect \ntrue differences in consciousness change dynamics. Second, our subject -level statistical \napproach properly accounts for repeated measures dependencie s, avoiding the \npseudoreplication problem that has undermined previous studies attempting to compare \nindividual transition events (Lazic, 2010). Third, our large sample of transitions (622 total) \nprovides adequate power to detect effects even with conservative statistical approaches. \nSeveral limitations warrant consideration. First, variance decomposition revealed that 96.6% \nof N1→N2 transition variance was within -subject rather than between -subject, reflecting the \nstochastic nature of spindle onset timing. While this high within-subject variability is consistent \nwith probabilistic threshold mechanisms in neural systems, it does limit precision of individual \ntransition predictions. Second, the REM→Wake analysis was based on only 32 transitions from \n7 of 12 subjects, reflecting the relat ive rarity of spontaneous REM awakenings in healthy \nsleepers. While the positive slope direction was consistent, additional data would strengthen \nconfidence in the quantitative estimate. Third, our sample consisted entirely of healthy young \nadults (age 25-35); generalizability to clinical populations with sleep disorders, older adults, or \ndeveloping children requires further investigation. Future studies should employ extended \nrecording sessions, targeted awakening protocols, and diverse populations to addr ess these \nlimitations. \nImplications for Consciousness Theory \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted November 11, 2025. ; https://doi.org/10.1101/2025.11.10.687628doi: bioRxiv preprint \n\nThe dual architecture model has profound implications for understanding consciousness \nmechanisms. Rather than representing mutually exclusive theories, IIT and GWT appear to \ndescribe complementary aspects of a unified system. IIT correctly predicts conscio usness \nchanges that emerge from continuous modulation of information integration capacity —these \nrepresent the majority of natural consciousness transitions. GWT correctly identifies discrete \nthreshold mechanisms mediated by specific neural circuits —these r epresent a minority but \ncritical class of transitions where specific gating functions are required. \nThis synthesis suggests that consciousness systems employ multiple control mechanisms \noptimized for different functional requirements. Gradual integration mechanisms allow fine -\ngrained modulation of consciousness level in response to changing demands. Threshold gating \nmechanisms provide discrete state boundaries that stabilize consciousness states against noise \nand prevent oscillation between states. The sleep cycle requires both: gradual mechanisms for \nsmooth transitions between most states, and a discrete  gate (spindle emergence) to reliably \nconsolidate Stage 2 sleep despite varying arousal pressure. \nOur findings predict that other consciousness -altering contexts (anesthesia, meditation, \npsychedelic states) will similarly exhibit both gradual and threshold transitions depending on \nthe specific neural mechanisms engaged. Testing this prediction across d iverse contexts \nrepresents an important avenue for future research. \nConclusion \nWe demonstrate that consciousness transitions follow a dual architecture combining gradual \nintegration mechanisms (three of four transition types, 75%) with discrete threshold gates (one \nof four transition types, 25%, specifically N1→N2 via spindle emergen ce). This reconciles \nIntegrated Information Theory and Global Workspace Theory by showing that both correctly \ndescribe consciousness change mechanisms operating in different contexts. The adaptive CGI \nframework successfully captures both transition types within a unified mathematical structure, \nproviding a quantitative tool for consciousness measurement across diverse contexts. Future \nwork should extend this framework to other consciousness-altering conditions and investigate \nwhether the 75/25 gradient -to-threshold ratio in transition types represents a fundamental \narchitectural principle of biological consciousness systems. \nReferences \nBerry, R. B., Brooks, R., Gamaldo, C. E., Harding, S. M., Marcus, C. L., & Vaughn, B. V. (2012). The \nAASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology and \nTechnical Specifications, Version 2.0. American Academy of Sleep Medicine. \nBrown, R. E., Basheer, R., McKenna, J. T., Strecker, R. E., & McCarley, R. W. (2012). Control of \nsleep and wakefulness. Physiological Reviews, 92(3), 1087-1187. \nCasali, A. G., Gosseries, O., Rosanova, M., Boly, M., Sarasso, S., Casali, K. R., ... & Massimini, M. \n(2013). A theoretically based index of consciousness independent of sensory processing and \nbehavior. Science Translational Medicine, 5(198), 198ra105. \nDang-Vu, T. T., McKinney, S. M., Buxton, O. M., Solet, J. M., & Ellenbogen, J. M. (2010). \nSpontaneous brain rhythms predict sleep stability in the face of noise. Current Biology, 20(15), \nR626-R627. \nDehaene, S., & Changeux, J. P. (2011). Experimental and theoretical approaches to conscious \nprocessing. Neuron, 70(2), 200-227. \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted November 11, 2025. ; https://doi.org/10.1101/2025.11.10.687628doi: bioRxiv preprint \n\nKemp, B., Zwinderman, A. H., Tuk, B., Kamphuisen, H. A., & Oberye, J. J. (2000). Analysis of a \nsleep-dependent neuronal feedback loop: the slow -wave microcontinuity of the EEG. IEEE \nTransactions on Biomedical Engineering, 47(9), 1185-1194. \nLazic, S. E. (2010). The problem of pseudoreplication in neuroscientific studies: is it affecting your \nanalysis? BMC Neuroscience, 11, 5. \nLüthi, A. (2014). Sleep spindles: where they come from, what they do. The Neuroscientist, 20(3), 243-\n256. \nPoulet, J. F., & Petersen, C. C. (2008). Internal brain state regulates membrane potential synchrony in \nbarrel cortex of behaving mice. Nature, 454(7206), 881-885. \nRechtschaffen, A., & Kales, A. (1968). A Manual of Standardized Terminology, Techniques and \nScoring System for Sleep Stages of Human Subjects. Washington, DC: US Government \nPrinting Office, US Public Health Service. \nTagliazucchi, E., Roseman, L., Kaelen, M., Orban, C., Muthukumaraswamy, S. D., Murphy, K., ... & \nCarhart-Harris, R. (2016). Increased global functional connectivity correlates with LSD -\ninduced ego dissolution. Current Biology, 26(8), 1043-1050. \nTononi, G., Boly, M., Massimini, M., & Koch, C. (2016). Integrated information theory: from \nconsciousness to its physical substrate. Nature Reviews Neuroscience, 17(7), 450-461. \nVyazovskiy, V. V., Olcese, U., Lazimy, Y. M., Faraguna, U., Esser, S. K., Williams, J. C., ... & Tononi, \nG. (2009). Cortical firing and sleep homeostasis. Neuron, 63(6), 865-878. \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted November 11, 2025. ; https://doi.org/10.1101/2025.11.10.687628doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}